Retrospective analysis of machine service logs and EPID‐based linac QA to investigate the effects of machine maintenance and calibration events on beam characteristics: A multi‐institutional study

BACKGROUND: Electronic Portal Imaging Devices (EPIDs) are widely used for LINAC quality assurance (QA) and underpin many manufacturer integrated quality control (MIQC) tools, yet the practical impact of routine linac service and calibration events on EPID-derived beam characteristics remains under-described. The Automated Quality Assurance (AQA) consortium has developed an independent, standardized EPID workflow for QA assessment that has been deployed across multiple TrueBeam linacs. We leveraged the AQA's framework and data to time-align vendor service logs with automated EPID analyses and characterize event-linked changes in photon beam output, symmetry, and collimator-jaw positioning. PURPOSE: To investigate how EPID-derived beam performance metrics change following common real-world service events, including beam output calibration, active beam steering, and collimator jaw position calibration. METHODS: We analyzed a multi-institutional dataset acquired between 2018 and 2024 across nine TrueBeam LINACs using the AQA Phase-2 EPID suite. EPID acquisitions were delivered in MV dosimetry mode and uploaded to a web portal for automated analysis with corresponding XML service logs to align service-event timelines with EPID QA results. Service events were grouped into beam-generation (dose calibration and active beam steering) and collimator (jaw position calibration) categories. The vendor XML service logs categorized events using labels such as "ANGLE R/T" or "POSITION servo" for beam-steering adjustments and "GAIN-OFFSET" for jaw position calibration. The EPID analysis included center-dose constancy, axial and transverse beam symmetry, collimator-centering correction, and X-/Y-jaw position. Measurement outliers were flagged using a density-based spatial clustering of applications with noise (DBSCAN) algorithm applied separately to one-dimensional EPID time series. RESULTS: Across seven linacs with documented 6 MV dose calibrations, EPID center-dose constancy showed a consistent post-calibration decrease relative to baseline (mean ± SD = -1.10% ± 0.26%; range: -0.73% to -1.54%). In a paired subset (two linacs; n = 12 calibration events), the difference between EPID-derived and ADCL-calibrated ion-chamber percent dose changes (ΔEPID - ΔIC) had a mean bias of 0.09% ± 0.79% (range -1.49% to 1.23%). Applying DBSCAN to the change in center-dose output identified outlier points that were associated with lower (negative) changes across all nine linacs. Following ANGLE R/T steering, EPID-derived symmetry shifts were observed and were concordant with daily detector-array readings in a representative case, while symmetry changes also occurred after non-steering activities (e.g., carousel/ion-chamber work), indicating multiple service-linked pathways affecting profile constancy. DBSCAN applied to jaw-position metrics flagged outliers exceeding routine variability (e.g., 0.44 mm vs 0.03 mm), including a documented 1.3 mm Y2 shift temporally linked to a jaw GAIN-OFFSET calibration. CONCLUSIONS: Integrating service logs with automated EPID analysis enabled event-level interpretation of longitudinal QA measurements. EPID center-dose changes following 6 MV dose calibrations were independently compared with ADCL-traceable ion-chamber measurements for a limited subset of two linacs, whereas the symmetry, jaw-position, and DBSCAN findings represent EPID-internal associations with documented service events and were not independently verified. Standardized EPID calibration records and prospectively timed pre- and post-service measurements are needed to strengthen attribution and confirm the broader applicability of these findings.